Anthropic / skill-creator
How do you write a skill that the agent actually loads?
The hard part is not the instructions. It is the description that decides when they load.
What to remember
- The description is the trigger. It matters more than the body.
- Skills should encode procedure, not general knowledge the model already has.
- Write from a workflow you have actually repeated, or you will encode an imagined process.
- Fewer, sharper skills beat a large library nobody can predict.
01What it changes in practice
A skill is only useful if it loads at the right moment. Most first attempts fail there: the instructions are fine, but the description is written in the author's vocabulary, so it never matches how anyone else describes the task.
Working with a skill that knows the format also stops the common mistake of writing an essay. A skill is a procedure with decision points, not documentation of a topic.
02Where it disappoints
It cannot tell you whether the workflow is worth encoding. That judgement comes from having done the work repeatedly and noticed the same steps.
Large skill libraries also become their own problem: overlapping descriptions make triggering unpredictable, and the fix is deletion rather than more skills.
Install and first run
- Install it from anthropics/skills under Apache 2.0.
- Start from a workflow you have already done three times by hand.
- Test the trigger by describing the task in the words a colleague would use, not the words you used when writing it.
Questions people actually ask
- What makes a skill trigger reliably?
- A description written in the words a user would use for the task, covering the phrasings you expect in practice.
- What should not go in a skill?
- General knowledge the model already has. Skills are for procedure, conventions and house specifics.
- How many skills is too many?
- When descriptions start overlapping, triggering becomes unpredictable. That is the signal to merge or delete.
Sources
Machine readable versions of this page